Google DeepMind and a group of philanthropic and public partners are moving quickly to address a problem they say is coming just as fast as new AI products: what happens when millions of autonomous AI agents start interacting with one another across the internet.

Early concerns emerge

In mid-June 2026, reporting highlighted that Google DeepMind is “worried about what happens when millions of agents start to interact,” as task‑performing AI systems are deployed at scale online.¹ Rohin Shah, who directs the firm’s AGI safety and alignment research, has framed this as a “whole new class of risk,” involving agents that can act without human oversight and even take instructions from other agents.¹

Shah argues that multi‑agent safety is underdeveloped as an academic field and that risks could escalate quickly once such systems spread through the economy.¹ The concern is that, like human institutions, large constellations of AI agents could display powerful collective behavior that no single system was explicitly designed to produce.¹

Funding call announced

Shortly after these concerns were aired, Google DeepMind, together with Schmidt Sciences, the Cooperative AI Foundation, the UK’s Advanced Research and Invention Agency (ARIA), and supported by Google.org, formally announced a “technical research funding call of up to $10M for researchers worldwide.”²

In its announcement, DeepMind said the world is “entering a new era” in which “millions of AI agents — built by different organizations — will interact across digital environments, communicating, negotiating and transacting with one another,” and stressed that these interactions must remain “safe and predictably.”²

Differing emphases, shared goal

From the human‑focused reporting side, the initiative is cast as an urgent attempt to “kick-start research outside tech companies” because “there just isn’t really a field of research for multi-agent safety yet.”¹ From DeepMind’s own perspective, the emphasis is on a “vital opportunity” to design frameworks that can “understand and mitigate against potential risks” and tackle “invisible safety risks” arising when independent systems interact across networks.²

Both perspectives converge on the same point: the behavior of large groups of AI agents, and the emergent effects they may generate, are poorly understood and require rapid, large‑scale research before these systems become deeply embedded in the global economy.¹²